A Multidisciplinary Survey and Framework for Design and Evaluation of Explainable AI Systems

Analysis

Why This Paper Matters

As AI systems become increasingly prevalent in everyday life, the need for interpretable and accountable intelligent systems has grown. Explainable AI (XAI) aims to enable systems to self-explain their reasoning, but researchers from different disciplines—machine learning, visualization, and human-computer interaction—often focus on different objectives and topics, leading to fragmented knowledge and challenges in consolidating design and evaluation methodologies. This paper addresses this gap by providing a multidisciplinary survey and framework that bridges these disciplines, offering a shared understanding of XAI design goals and evaluation methods.

The significance of this work lies in its effort to unify diverse perspectives and provide practical guidance for XAI teams. By categorizing design goals and evaluation methods and mapping them to different user groups, the authors enable teams to make informed decisions about how to design and evaluate explainable systems. This is particularly important as XAI research continues to expand rapidly, and practitioners need structured approaches to navigate the complex landscape of methods and goals.

Technical Contributions

• A comprehensive review of XAI literature across machine learning, visualization, and HCI, resulting in a categorization of design goals and evaluation methods.
• A mapping between design goals for different XAI user groups (e.g., developers, end-users, domain experts) and appropriate evaluation methods.
• A step-by-step design and evaluation framework that guides multidisciplinary teams through iterative cycles, from goal definition to evaluation.
• Ready-to-use tables summarizing evaluation methods and recommendations for various XAI goals, facilitating practical application.

Results

The paper does not present empirical results from a user study or experiment; instead, it offers a synthesis of existing literature. The main outcome is the categorization and framework, which are derived from a thorough review of XAI papers. The authors provide summarized tables that list evaluation methods (e.g., user studies, quantitative metrics, expert reviews) and recommendations for different design goals (e.g., trust, understandability, usability). These tables serve as a practical resource for XAI practitioners.

Significance

This survey and framework have significant implications for the AI field, particularly for interdisciplinary XAI research and practice. By consolidating knowledge across disciplines, the paper helps researchers and practitioners avoid reinventing the wheel and promotes the adoption of appropriate evaluation methods. The framework encourages systematic design and evaluation of XAI systems, which can lead to more effective and trustworthy AI applications. With 781 citations, this work has already influenced the XAI community and continues to serve as a foundational reference for designing and evaluating explainable systems.